Missing Physiologic Data in NTDB and TQP: How Much Complete-Case Analysis Actually Biases Trauma Mortality

Published September 20, 2026

Complete-case analysis is the quiet default in trauma registry research. Drop the rows with a missing GCS, blood pressure, or respiratory rate, fit the model on whatever is left, and report the result. A 2018 systematic review of 539 trauma registry manuscripts found that more than half either never mentioned data completeness or mentioned missing data without saying how it was handled, which means complete-case analysis by default, undocumented, is closer to the norm than the exception. We built an independent simulation, run through two statistical engines, to put an exact number on what that default actually costs.

Why This Missingness Is Not Random

Emergency department GCS, systolic blood pressure, and respiratory rate go missing because a clinician could not fully assess the patient, and the patients hardest to assess are disproportionately the ones who arrived intubated, sedated, or in extremis. An analysis of 2,520 cases from the Victorian State Trauma Registry confirmed the pattern directly, finding that respiratory rate, GCS, an intubation qualifier flag, and systolic blood pressure carried the greatest share of missing observations of any variable needed for standard trauma risk adjustment, and that a clinical logic, not chance, predicted which patients had the gaps. Statisticians call this missing not at random, since the probability that a value is missing depends on the value itself, low GCS and unstable vitals get dropped from the chart at exactly the rate that severe injury drives sedation and intubation. Complete-case analysis silently assumes the opposite.

The Simulation

We generated a synthetic cohort of 10,000 trauma admissions per replicate, with two partially correlated latent severity components, one driving anatomic injury severity and hemodynamic status (ISS, systolic blood pressure, respiratory rate), the other driving neurologic status (GCS), consistent with the moderate, not perfect, correlation between anatomic and neurologic injury seen in real polytrauma. True mortality was generated from a logistic model incorporating age, mechanism, ISS, GCS, systolic blood pressure, and respiratory rate. Missingness on GCS, systolic blood pressure, and respiratory rate was then imposed exactly as the literature describes it, as a function of the patient’s own true physiologic value plus an intubation indicator, so that missingness rates landed in a realistic range, roughly 30% for GCS, 24% for systolic blood pressure, and 16% for respiratory rate, comparable in order of magnitude to the missingness proportions reported for these same three variables in published trauma registry audits.

Four analytic approaches were then run on every replicate. An oracle model fit on the true, fully observed physiologic values, standing in for the analysis a researcher would run if nothing were missing. A complete-case model, dropping every patient missing any of the three variables, mirroring undocumented default practice. A naive single mean-imputation model, replacing each missing value with that variable’s overall mean, a shortcut still seen in submitted manuscripts. And a proper multiple imputation model, following the approach the NTDB-specific literature recommends, with the mortality outcome itself included in the imputation model and results pooled across imputations by Rubin’s rules. The whole pipeline was coded independently in Python and in R, using different imputation implementations (scikit-learn’s iterative imputer against R’s mice package), and run for 300 replicates in Python and 100 in R. Agreement between the two, on both the direction and the approximate size of every effect below, is what certifies the result rather than a single script’s idiosyncrasies.

Finding One: Complete-Case Analysis Roughly Halves the Reported Mortality Rate

True population mortality across the simulated cohort averaged 5.50%. Restricting the analysis to patients with complete GCS, systolic blood pressure, and respiratory rate, exactly what an undocumented complete-case analysis does, dropped the observed mortality rate to 2.68%, a reduction of more than half. A mechanical process, not a statistical one, produces the gap. Excluding a patient from the analysis because they were too unstable to have a full physiologic assessment recorded is exactly equivalent to excluding the patients most likely to die, and no amount of downstream modelling sophistication corrects a mortality rate computed on a systematically healthier subgroup than the one actually treated. A benchmarking study or quality report that quotes a raw or lightly adjusted mortality rate from a complete-case trauma cohort is, on this evidence, reporting a number that understates true mortality by a wide margin, not a rounding-level discrepancy.

Our own result echoes what a Monte Carlo simulation built on 63,020 real NTDB patients across 68 hospitals found for a related but distinct question, whether the choice of missing-data method changes which hospitals get labelled a quality outlier on observed-to-expected mortality. That earlier work varied the analytic method while holding the underlying cohort fixed. Ours varies the underlying missingness mechanism itself, synthetically, so that the true value being estimated is always known, which is what makes it possible to state an exact bias rather than only a rank-order disagreement between methods.

Finding Two: The Adjusted Effect Survives Complete-Case Restriction Better Than Expected

The risk-adjusted effect of GCS on mortality, expressed as the model’s log-odds coefficient, averaged -0.300 in the oracle model. In the complete-case model, restricted to the roughly half of patients with all three variables observed, the coefficient averaged -0.301, close enough to the oracle value that the loss is better described as reduced precision from a smaller, more homogeneous sample than as directional bias. A specific, useful asymmetry sits behind that result. Harm from complete-case analysis in this setting concentrates in the marginal quantity, the reported mortality rate and case mix, rather than in the conditional, risk-adjusted coefficient that most trauma papers actually report as their headline finding. A paper’s adjusted odds ratio for GCS may be approximately right even while its stated cohort mortality rate is badly wrong, which is exactly the kind of selective damage a reviewer scanning only the abstract’s headline statistic will miss.

Finding Three: Naive Mean Imputation Attenuates the Adjusted Effect That Complete-Case Analysis Preserves

Replacing missing GCS, systolic blood pressure, and respiratory rate values with each variable’s overall mean, a shortcut that at least keeps every patient in the analysis, produced a GCS coefficient averaging -0.165, an attenuation of roughly half relative to the oracle value. A direct mechanism, not a subtle one, explains why. Filling every missing GCS with the same constant destroys precisely the variation in GCS that the regression needs to estimate its effect on mortality, and it does so hardest for exactly the patients whose true GCS was most extreme, since they are the ones most likely to be missing in the first place. Mean imputation trades the mortality-rate bias of complete-case analysis for a different, and in this simulation larger, bias in the adjusted effect that the study is actually built to estimate.

Finding Four: Multiple Imputation With the Outcome Included Recovers Most of the Loss

Multiple imputation, applied with the mortality outcome included in the imputation model and pooled across imputations by Rubin’s rules, produced a GCS coefficient averaging -0.263, recovering the large majority of the gap that naive mean imputation left behind, while also keeping every patient in the mortality-rate calculation rather than dropping the sickest ones as complete-case analysis does. This is the same design choice, outcome-inclusive multiple imputation, that the NTDB-specific methodological literature has recommended since at least 2009, and the simulation here reproduces the reason that recommendation holds independently of any single dataset, using entirely synthetic data built to isolate the mechanism rather than any real registry’s idiosyncrasies.

Applying This to Your Own Registry Extract

Four checks translate this simulation into practice on a real participant use file. Quantify the missingness rate for GCS, systolic blood pressure, and respiratory rate separately, rather than reporting a single combined completeness figure that hides which variable is driving the problem. Test whether missingness on each variable is associated with other markers of severity that are fully observed, an intubation flag, arrival by helicopter, ISS itself, since a clear association is the practical signature of the informative missingness this simulation models. Run multiple imputation with the mortality outcome, and any other outcome the study reports, included as a predictor in the imputation model, not left out on the assumption that outcomes belong only in the analysis stage. And report the complete-case result alongside the multiple-imputation result rather than in place of it, so a reviewer can see the size of the correction directly instead of taking it on faith.

Free Tools for This Kind of Check

• Six techniques for cleaning clinical registry data, for the missingness audit that should precede any imputation decision
• STROBE reporting checklist tutorial, for how a missing-data flow should be documented in the manuscript itself
• NTDB and TQP database explainer, for the wider set of design issues beyond missingness that a TQP study needs to address

What This Means for a Manuscript

Three practical conclusions follow directly. A stated mortality rate or case-fatality proportion drawn from a trauma registry cohort deserves specific scrutiny of how the denominator was built, since a complete-case restriction can plausibly cut a reported rate by half without any error message or obvious red flag in the output. An adjusted odds ratio for a physiologic covariate is comparatively more robust to complete-case handling than the raw rate is, though reduced sample size still costs precision and widens confidence intervals. And of the two common shortcuts, dropping incomplete cases or filling gaps with the sample mean, multiple imputation with the outcome variable included beats both, which is exactly the practice Editverse’s own NTDB and TQP Trauma Registry Study methodology commits to for every engagement it runs.

Results Summary Across 300 Python and 100 R Replicates

Only complete-case analysis changes the reported mortality rate, since it is the only method that changes which patients are analysed rather than filling in what is missing about them. Only the two imputation-based methods keep the full cohort in the mortality calculation, and of those two, outcome-inclusive multiple imputation is the one that also protects the risk-adjusted effect size.

Limits of This Simulation

This is a synthetic illustration built to isolate one mechanism cleanly, not a re-analysis of real TQP data, since Editverse does not hold a TQP participant use file agreement and does not obtain the dataset on a client’s behalf. The missingness rates, correlation structure, and mortality model were chosen to sit in a realistic range based on the published literature cited throughout, not fitted to a specific registry year. A client engagement works from the client’s own extract, characterises that file’s actual missingness pattern directly, and reports a complete-case sensitivity analysis alongside the primary multiple-imputation result exactly as described in our NTDB and TQP explainer.

Professional Services

Editverse builds NTDB and TQP trauma outcome studies with physiologic missingness handled by outcome-inclusive multiple imputation and a complete-case sensitivity analysis reported alongside it, exactly as validated here.

References

1. Shivasabesan G, Mitra B, O’Reilly GM. Missing data in trauma registries: A systematic review. Injury. 2018;49(9):1641-1647. doi: 10.1016/j.injury.2018.03.035
2. O’Reilly GM, Cameron PA, Jolley DJ. Which patients have missing data? An analysis of missingness in a trauma registry. Injury. 2013;43(11):1917-1923. doi: 10.1016/j.injury.2012.07.185
3. Moore L, Hanley JA, Turgeon AF, Lavoie A, Emond M. A multiple imputation model for imputing missing physiologic data in the national trauma data bank. Journal of the American College of Surgeons. 2009;209(5):572-579. doi: 10.1016/j.jamcollsurg.2009.07.004
4. Glance LG, Osler TM, Mukamel DB, Meredith W, Dick AW. Impact of statistical approaches for handling missing data on trauma center quality. Annals of Surgery. 2009;249(1):143-148. doi: 10.1097/SLA.0b013e31818e544b
5. Newgard CD. The validity of using multiple imputation for missing out-of-hospital data in a state trauma registry. Academic Emergency Medicine. 2006;13(3):314-324. doi: 10.1197/j.aem.2005.09.011
6. O’Reilly GM, Jolley DJ, Cameron PA, Gabbe B. Missing in action: a case study of the application of methods for dealing with missing data to trauma system benchmarking. Academic Emergency Medicine. 2011;17(10):1122-1129. doi: 10.1111/j.1553-2712.2010.00887.x